Research On Automatic Defrost Control Strategy Based On Intelligent Algorithm

Frequent switching of the door of the cold storage delivers a great deal of hot and humid air from outside into the cold storage.There is also a large amount of heat and moisture exchange in the process of cooling and preservation of goods.Water molecules carried by wet air will condense and change from vaper into frost when passing through the fan and fintube heat exchanger.With the accumulation of frost,the heat transfer capacity of heat exchanger decreases gradually,which means that the system needs to run longer and consume more energy to reach the same temperature.In order to prevent this phenomenon,the most commonly used method is to defrost regularly.For those cold storages with higher operation and management level,the system is shut down and defrosting just because of this timing mechanism,even if defrosting is not necessary in some cases.The increase in energy consumption of cold storage is a waste.Therefore,it is particularly important to study the starting point of defrosting and the duration of defrosting,in order to achieve ‘on-demand defrosting'.In this paper,the defrosting control strategy of finned-tube heat exchanger is studied,both experimentally and modeled.It is difficult to measure frost amount by weighing the weight of defrosting water in a short time accurately,the method of calculating the cumulative mass of moisture to approximate the frost amount would be a common option.Firstly,considering the idea of "infinitesimal element method",the frosting process,or more precisely the dehumidifying process,is transformed into the superposition of the process of moisture precipitation from wet air in every single minute.The higher accuracy of the calculation of moisture precipitation,the better simulation to the mass of frost.When the cumulation reaches the threshold,the system can send a defrosting signal and perform a defrosting operation.Then,a lumped parameter model of frosting process of fin-tube heat exchanger is established following the law of conservation of energy and mass,which includes two parts: heat transfer model and frost growth model.Under specific working conditions,the heat transfer of wet air passing through the control unit of the heat exchanger is calculated by logarithmic mean temperature difference method and the dichotomy method.Then,the moisture precipitation of wet air in every single minute can be calculated according to the moist air parameters at inlet,and the cumulative results of moisture precipitation in a period of time are given.At the same time,the theory and mathematical expression of BP neural network are introduced,together with its application in the field of refrigeration system.Models to solve the starting point of defrosting and defrosting duration are established respectively by the mass of frost prediction model and the defrosting duration prediction model,both using BP neural network.The training process of the model is presented in the flow chart,and the evaluation of the model,indicators and necessary optimizations are proposed as well.Thirdly,the low temperature wind tunnel test bench is introduced,including the temperature and humidity control system,the testing system,the data acquisition system,etc.,the heat exchange capacity of fin-tube heat exchanger which is used to carry out the experiment is calculated and checked too.Subsequently,the frosting process of the fin-andtube heat exchanger under different working conditions and different air temperature,relative humidity and wind speed was studied experimentally.During the experiment,the parameters such as temperature,relative humidity,air volume,and moisture content of the wet air were collected.The duration of the experiment and the duration of the defrosting were also recorded to make preparation for training the neural network.Finally,the experimental data after cleaning is used to determine the number of hidden nodes of both network and then to train and test two models,respectively.The performance of the network is analyzed by the performance curve of the training process,the regression analysis between the targets and output of neural network,the statistics methods,and the comparison of the testing data.The matrix of weights of the neural network is given under the optimal solution.According to the characteristics of the two networks,the optimization direction is proposed separately.The best neural network models in this paper have the following results:(1)As the training process progresses,the performance curves of the two neural network models both show a downward trend,indicating that the model is gradually converging.The mass of frost prediction model converges after the 25385 th training.The mean square error is 0.00070588,which is less than the training set value of 0.01,and the anti-normalized value is 5.138,indicating that the mean square error between the targets and outputs of the neural network is 5.138 grams.Its standard deviation accounts for approximately 11.21% of the mean weight of moisture precipitation.The defrosting duration prediction model converges after the 139 th training.The mean square error value at this moment is 0.0011469,which is less than the training set value of 0.005.The value is anti-normalized to 3.992,indicating that the mean square error between the targets and outputs of the neural network is 3.992 minutes.Its standard deviation accounts for approximately 3.666% of the mean weight of moisture precipitation.(2)In the mass of frost prediction model,the regression coefficients R between the targets and outputs of the neural network on the training dataset,the validation dataset and the testing dataset are 0.96125,0.97197,and 0.96443,respectively,both exceeding 96%.Among the 1062 samples used for testing,the data with the error within 10% is 758,accounting for 71.37%;those within 20% is 979,accounting for 92.18%;the average error of all test data is 10.1061%.In the defrosting duration prediction model,on the training dataset,the validation dataset,the testing dataset and all data set,the regression coefficients R are 0.99158,0.99783,0.98618,and 0.99077,respectively,both exceeding 98%.The error of the four testing sets are 0.0047%?-3.59%?-7% and-4.21% respectively,acceptable.Based on the above indicators,it can be concluded that the prediction accuracy of the two neural network models proposed in this paper is acceptable and can be used to predict frost amount and defrosting time.